# cl4ud1a — LoRA adapter (Z-Image Turbo, rank=16)

This folder holds the training experiment named `cl4ud1a` — a LoRA adapter for Tongyi-MAI/Z-Image-Turbo (Z-Image Turbo) tuned with rank 16 parameters.

Key facts
- Base model: Tongyi-MAI/Z-Image-Turbo (arch: zimage:turbo)
- LoRA type: UNet LoRA (linear=16, conv=16, alpha=16)
- Training steps: 3000 (checkpoints saved every 250 steps)
- Save format: Diffusers safetensors (dtype: bf16)
- Training device: cuda
- Quantization: qfloat8 applied to model and text encoder

Artifacts in this folder
- `cl4ud1a.safetensors` — final LoRA/adapted weights (merged in training pipeline)
- `cl4ud1a_00000XXXXX.safetensors` — saved checkpoints (step increments)
- `optimizer.pt` — optimizer state (checkpoint)
- `config.yaml` — original run configuration used for training
- `log.txt` — raw training log (progress, warnings, reproducibility notes)
- `samples/` — generated sample images at a few checkpoints

Training & notes observed
- Training used a small dataset (~14 images, augmented to 28 via flips) at mixed resolutions (768–1024). Latents and text embeddings were cached for speed.
- A PIL-based EXIF parsing error appeared for one PNG during preprocessing; dataset sanitation is recommended before reproduction (see log snippet).
- Assistant LoRA adapter was loaded/merged during training — see `config.yaml` for assistant adapter path.

How to reproduce (short)
1. Ensure you have the same base model (Tongyi-MAI/Z-Image-Turbo) accessible.
2. Recreate the environment with GPU + CUDA and BF16 support.
3. Use `config.yaml` to re-run the trainer used by the author (dataset paths will need adjustment).

Usage example (consumer)
- To apply the LoRA at inference time, use your Z-Image-Turbo-compatible pipeline loader and merge or inject the safetensors file into the UNet weights (example depends on your runner/adapter).

If you plan to upload this experiment to Hugging Face: include `cl4ud1a.safetensors`, `config.yaml`, `log.txt` and a short model card describing license and data provenance.